The Pangloss of Pangram

On "detection" of AI text, and whether we should care.

A pangram is a sentence that uses every letter of the alphabet. The most famous example: the quick brown fox jumps over the lazy dog. Pangrams focus on the components of a sentence, rather than the message it produces. Wordplay like this is entertaining, and, if not that, then harmless. Unless there’s a pangloss applied to the pangram. Dr. Pangloss, the tutor in Voltaire’s 1759 satire Candide, has one proclamation for everything: it’s the best of all possible worlds. No wreckage changes his mind, because the verdict arrived before the evidence.

Last month, Substack switched on AI detection, partnering with a company called Pangram. Readers can scan a post over 100 words and get a score on its composition: human, AI-assisted, or AI. There’s no lookup involved. No database. No measuring the predictability of your words. It’s a pattern recognizer trained on millions of documents by people and by machines. It reads word choice, syntax, the shape of a sentence. Never what the sentence is for. To its credit, the machine shows a spectrum. But its readers rarely do. It becomes a scan-and-shame ritual, with any hint of AI rounding to a single judgment. And so, Pangloss is inverted. He was sure of the best of all possible worlds. They’re sure of the worst of all possible words.

A reader with Pangram gets an estimate of where words came from, but not a ruling on whether they’re worth reading. Substack’s own announcement admits that ceiling: Pangram “can only detect whether AI was used to make the text, not whether great human care went into creating it.” Substack CEO Chris Best adds: AI use isn’t the problem, and AI didn’t invent slop. Yet people read the score as a ruling anyway, and revive a trial humans have run for centuries.

The morning after Lincoln spoke at Gettysburg, the Chicago Times reviewed not the address but the man: “The cheek of every American must tingle with shame as he reads the silly, flat and dishwatery utterances of the man who has to be pointed out to intelligent foreigners as the President of the United States.” The verdicts split neatly by party; every paper was scanning the speaker. It took Harrisburg’s paper 150 years to retract its sneer: “the Patriot-News regrets the error.”

Eventually the attack refined itself: don’t call writers silly; ask whether the words are theirs. It found Martin Luther King Jr. too, decades after his death. Scholars had traced what his “I Have a Dream” speech owed to a 1952 convention address, to a preacher’s refrain, and to the hymnal, and they called it tradition. Detractors read the same list as an indictment. The tally changed nothing. Nobody hands that speech back.

But consider what the tally actually turns up. By the account of the late Clarence Jones, King’s attorney and a co-drafter of the speech, at least nine people shaped the prepared text. Does nine diminish it? Would fifty? Would a thousand? Somewhere in that climb the question is supposed to flip the verdict, and nobody can say where. King stood at that podium and meant every word he said, and the counting never mattered. The same arithmetic gets run on anyone who drafts with a model, as if the answer were a headcount. Would we hand those words back?

I hope not. We ran that play on a wartime president and a preacher remaking a nation. The words won anyway. Now we run it on each other’s newsletters. Same stones, less excuse. And, no, the small stage doesn’t lower the stakes. Nobody knows which piece of writing has an idea that matters, so our choice is simple: tear at each other, or tease out the ideas worth holding.

This returns me to the detector. Where do I stand? Imagine a spectrum from 0 to 10. Zero is the pitchfork crowd: AI is plagiarism and its practitioners are charlatans. Ten is the full surrender: let the software run and leave the room. I land near 8. I’ve spent years making the case that these systems enhance our expression and depth of thought. But, as any sensible person avoids the self-ridiculing extremes, I stop well short of 10.

Examine that murky middle. What should “AI-assisted” cover? A tighter verb the software suggested? Your own spoken words, transcribed by machine? The label doesn’t say, and by the platform’s own admission the detector can’t fully sort them—it can’t even tell when AI was merely the source. But readers won’t parse any of that. They’ll see “assisted” and fill in whatever they already suspected.

Ultimately, an AI detection score is a Rorschach test. It reflects the reader’s existing biases about the tech rather than the quality of the prose. The scan claims to examine the writing, but it’s the scanner who ends up examined. Tell me how you react to a 70% human score from Pangram, and I’ll tell you what you already believed about AI. And still nothing about whether the piece was worth your time.

Today, the absurdity is visible from the street. The labels are blunt, and the ritual is loud. But wait a year. The models will gain polish. The detectors will learn to hedge. And the underlying error will be hidden, not solved. Laugh now, while all this is still obvious.

None of this lets writers off the hook. There’s still a promise to keep. My name on a page is a vow. I invested energy in these thoughts. I am present in these words. That vow never required typing every letter manually. We didn’t ask that of the novelist who dictates or the leader who has a ghostwriter. The true transgression is different: hitting a button, not bothering to weigh or even read the output, and publishing it as your own. Regardless of the tools used, a human being must mean what is said.

If that makes me the “LinkedIn guy,” excellent. This label uses LinkedIn as shorthand for slop. Boring, promotional, maybe. It gets people paid. I’ll take that over pure and useless. I’m for AI when it empowers people—which is the only reason to build it.

Prize the work, not my quill, keys, or vexing bot; judge craft. Some of you will have noticed that the last line was a pangram or checked if AI drafted it, but those are the least interesting things about it.